Non-equidistant integration for gMAM

[oameye]

The current bottleneck in gMAM is the interpolation step which has to be doen every optimization step. This also makes it that any other solver than gradient descent and LBGFS is not usable in the current implementation with Optim.jl. The reason why we have to interpolate is to make the path equidistant to allow for easy integration needed for the action. Is this the only reason? In case so we could opt for non-equidistant integration to compute the action.